A University of Wollongong lecturer has pitted students studying criminal law against ten AI-generated responses in their 2023 final exam, as part of his study into the capabilities of AI.
A criminal law lecturer at the University of Wollongong, Dr Armin Alimardani, secretly mixed the AI responses with student responses and examined the difference in exam results, to find that only two AI-generated responses scored higher than the student average.
Dr Alimardani said that the results of the AI responses in this study failed to meet a claim by OpenAI, that AI could outperform 90 per cent of students in a US bar exam.
“They were nowhere near close to that benchmark of the OpenAI release,” he said.
“We have got to be careful with how we interpret these kinds of announcements by AI companies.”
Dr Alimardani said that whilst the AI responses performed well on simple questions, they struggled on questions involving legal analysis.
“AI models are good at essay writing and short answer questions because they are not that complicated,” he said.
“The main problem was that they didn’t perform critical analysis or legal analysis at the level that we expect from high achieving or high ranked students.”
Video: Dr Armin Alimardani explaining the results of his study into the comparison of AI and law students
Dr Alimardani said that despite the AI models not performing well, none of the exam markers, who were all unaware of the involvement of AI, flagged the papers for containing AI-generated text.
“One other thing that surprised me was that none of the teachers who marked the AI papers, even suspected that there was something wrong with those papers,” he said.
“This suggests that we are past the line that we can identify AI-generated text. Sometimes we can guess, but it’s very difficult at this time to say whether something is AI-generated or not.”
Juliana Peloche, an AI researcher and professor at the University of Wollongong’s School of Education, mentioned that while we can still differentiate between AI-generated and human-written text for now, this distinction may not last much longer.
“We are not there yet, but every day we are getting more data into the systems and are heading to this point where we won’t be able to tell the difference between an AI-generated text and a human-generated text.”
Nicholas Ryan, a University of Wollongong student involved in Dr Alimardani’s study, stated that the research highlights how AI will increasingly become a significant part of our future reality.
“To be part of a study like this is kind of interesting because I think it makes it feel a little bit more real,” he said.
“Most people have this general sort of anticipation towards AI and the more you’re exposed to it and there more real studies and real applications, you realise that it is coming very shortly and it’s going to be integrated within our life soon.”
However, Ms Peloche said that as AI becomes part of reality, it also requires regulation.
“We are at the moment, what we call a wild west because there are no regulations in AI,” she said.
“I think there should be a little bit more transparency from the big companies, but the legislation should actually come from the government.”
This comes a few days after a Melbourne lawyer was referred to the Victorian Legal Complaint’s body for using AI-generated citations in the Family Court.
